Download Evaluation of water using variables

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Particle-size distribution wikipedia , lookup

Transcript
ENVlRONMETRlCS
Envimnmdncs 2003: 14: 373-386 (DOI: IO.IOWenv.592)
Evaluation of water quality using acceptance sampling by variables
Eric P. ~ m i t h * . t ,Alyaa Zahran, Mahmoud Mahmoud and Keying Ye
Doparlmonl of
Slolislics. Virginia Tech. Blocksburg. VA 24061-0439, U.SA
SUMMARY
Under section 303(d) of the Clean Water Act, states must identify water segments where loads of. pollutants are
violating numeric water quality standards. Consequences of misidentification are quite important. A decision that
water quality is impaired initiates the total maximum daily load or TMDL planning requirement. Falsely
concludin~that a water segment is impaired results in unnecessary TMDL plannine and oollution control
i~n~lemeniation
costs. On the other hand. hlscly concluding that a ,&ment is not i n l p ~ ~ r cm.8~
d pose a nrk to
human health or to the sen ices of the aquatic environment. Becausc of the cunsequencc,, a method 1s dc,ired that
minimizes or controls the error rates. The most commonlv. applied
aooroach
is to use the Environmental
..
..
Protection Agency (EPA)'s mw score approach in which a stream scgmrnr is lirted 31 impired when grcaler than
10 per cent of the measurements of water quality conditions exceed a numeric criteria. An alternative to the EPA
aovroach is the binomial test that the orooortion exceedine the standard is 0.10 or less. This ao~roachuses the
number of samples exceeding the criteha as a tesi statistic Gong with the binomial distribution fi; evaluation and
estimation of error rates. Both approaches treat measurements as binary; thevalues either exceed or d o not exceed
the standard. An alternative approach is to use thc actual numerical ;ulucr to eval~atcal~ndard.This method is
referred to as variables acceptance sampling in quality control literature. The methods are compared on the basis
of error rates.
If certain assumptions are met then the variables acceptance method is superior in the sense that the variables
acceptance method requires smaller sample sizes to achieve the same error rates as the raw score method or the
binomial method. Issues associated with potential problems with environmental measurements and adjustments
for their effects are discussed. Copyright 02003 John Wiley & Sons, Ltd.
~
KEY
~
WOKDS: TMDL; monitoring; standards; acceptance sampling by variables; binomial distribution
1. INTRODUCTION
I n the United States, every state is required under Section 303(d) o f the Clean Water Act to identify
water segments where anthropogenic pollution is leading to violations of water quality standards. A
standard is a numerical criterion that is set a s an indication of a healthy water body. Water segments
that are in violation are listed and may h e required to undergo a further assessment known as the Total
Maximum Daily Load (TMDL) process. T h e TMDL process is a study of the watershed near the site to
*Correspondence to: Eric P. Smith. Dcpilrtment of Statistics, Virginia Tech. Blacksburg. VA 24061.0439, U.S.A.
'E-mail: [email protected]
Contract/grant sponsor: U.S. Environmental Protection Agency's Science to Achieve Results (STAR) contract/grant number: R82795301. Canwcl/grant sponsor: Virginia Water Resources Research Center. Published online 10 April 2003
Copyright 02003 John Wiley & Sons, Ltd.
Received 5 October 2002
Revised I 2 April 2002
374
E. P. SMITH ETAL.
determine the amount of pollutant that may be tolerated by the watershed without violating the
standard. Part of the study involves identification of pollution sources, steps for reducing the pollution
load to the system and a plan for reducing pollution inputs. Clearly this can be a costly process and
state agencies have expressed concerns about the monitoring and listing process. For example, the
State of Virginia monitors over 17000 miles of watenvays that are divided into segments. A set of
segments within the waterways is sampled on a (typically) monthly or quarterly basis. Since the law
requires an assessment on a two-year cycle the monitoring program results in a large amount of
information. However, decisions are made on a site-by-site basis and the sample size for an individual
site is often small. For example, quarterly sampling typically results in eight observations to make the
decision to list or not list the segment.
Critical to the listing process are the data that are collected and the decision process used to list a
segment. The challenge to the assessor is to use the limited data that is available to determine if the
stream is violating standards, given that samples might be affected by variation and natural or
background conditions. The typical approach for making a decision involves both objective and
subjective methods. All information that is collected is to be used in the process, and this might include
anecdotal information, testimony, and citizen monitoring data as well as samples collected by the
agency. The objective approach involves a simple test of the data that we refer to as the 'raw score'
approach. Specifically if more than 10 per cent of the samples exceed the standard then the site is
declared to not meet the usability criteria.
From a statistical perspective, the conceptual model for evaluation of water quality is presented in
Figure 1. In Figure 1 the measurement is a concentration of some contaminant in the ambient water.
The distribution of the concentration represents a possible range of values that might be observed at a
site over the time period of interest and the standard is one possible concentration. The standard may
have been chosen based on laboratory data, previous field data or expert opinion. Suppose the water
quality guidelines require that a concentration of 8.0 or less should be met 90 per cent or more of
the time. In the raw score approach the proportion of samples that exceed the standard is calculated
and, if this is greater than 10 per cent, the site is listed. For example, if there are five samples and one
or more exceeds the standard, the site is declared impaired. The same is true for all sample sizes
between one and nine. For sample sizes between 10 and 19, one sample is allowed to exceed the
standard but not more. The diagram suggests several other approaches to the analysis of data arising
under this scheme.
0.25
0
5
standard
1
10
15
Concentration
Figure I. Hypothetical distributions of measuremenu for listing and not listing a site as impaired based an measurements
Copyright 0 2003 John Wiley & Sons, Ltd.
Envimnmerdcs 2003; 14: 373-386
EVALUATION OF WATER QUALITY
0I
0.~0
10
20
30
40
I
I
50
60
binomial
raw score
Sample size
Figure 2. Probability of listing a site that is not impaired (Type I errorrate) for different sample sizes forthe raw score method
and the binomial method choosing a cut point to bound the enox at 0.20
One alternative described in Smith et al. (2001) uses a binomial testing approach for evaluation of a
standard (see also McBride and Ellis, 2001, for a different approach). Smith et al. (2001) viewed the
assessment problem as a decision process and recognized that error rates should be considered.
Specifically, let p be the probability of exceeding the standard. Then the problem may be set up using
hypothesis testing with Ho:p<po (no impairment, don't list) versus HI: p > po (impairment, list).
The value ofpo would be set to 0.10 and the test based on j,the proportion of measurements exceeding
the standard.
The hypothesis testing scenario suggests evaluation of tests based on error rates. From the
environmental perspective, the assessor needs to be concerned about falsely listing a site and failing
to list a site that is in poor condition. Falsely listing a site may trigger the TMDL process and incur
unnecessw constraints on agriculture or industry. Failing to list a site that has poor water quality may
result in increased risk to human and ecological health. If error rates are considered, Smith et al.
(2001) show that the raw score method has a strong tendency to falsely list a site. The binomial method
as might be typically applied (i.e. using a Type I error rate of 0.05) will have a tendency to not list sites
that should be listed (Figures 2 and 3). However, the binomial approach allows for control of both
errors, and sampling plans may be developed to set error rates at satisfactory levels for sufficient
sample sizes. Both methods may be criticized, as the actual numerical value is not accounted for. In the
above example, a value of 8.1 is treated the same as a value of 801; it simply exceeds the standard of 8.
An alternative to these approaches is based on acceptance sampling by variables.
Acceptance sampling is a major field in statistical quality control where the view is that a group or
lot of manufactured items is to be accepted or rejected. Based on a sample from the lot, a manager
makes a decision. Evaluation of the method is based on the risk of accepting lots of given quality, and
sampling plans are developed to try to minimize error rates.
There are two approaches to acceptance sampling in the literature. The first approach is acceptance
sampling by attribute, in which the product i s specified as defective or non-defective based on a certain
cutoff point. The number of defectives is used as the test statistic. This is the same as the binomial test.
The other type is acceptance sampling by variables, where the value of the measurements is taken
into account when calculating the test statistic. There are two approaches to the problem. One is to
calculate, given the probability model, the area under the probability density that is more extreme than
the standard and evaluate if this exceeds the desired impairment probability (in our case 10 per cent).
Copyright 8 2003 John Wiley & Sons. Ltd.
Envimnnterrics 2003;14: 373-386
376
E. P. SMITH ETAL.
The second approach is to calculate the value of a parameter, typically the mean, that results in the
distribution having a standard with associated probability of 10 per cent. The mean of the observed
data is then compared with the derived mean. As the method uses actual values, it may be more
informative. This is reflected in the fact that acceptance sampling by variables requires a smaller
sample size than the attributes plans to have the same operating characteristic curve (OC), i.e. the
probability of accepting a lot of items as a function of the proportion of defectives. However, in
acceptance sampling by variables the distribution of the measurements must be known.
Since World Warn, variables-acceptance sampling plans have been in use, although primarily in an
industrial quality control context. Under the assumption that the characteristic being measured is
normally distributed, many plans have been developed. Lieberman and Resnikoff (1955) developed an
OC-matching collection of plans based on known variance, unknown variance and average range.
These plans were indexed by the acceptable quality level (AQL, i.e. producer's risk). Owen (1966)
introduced one-sided variables sampling plans that are indexed by the AQL as well as the lot tolerance
percent defective (LTPD, i.e. consumer's risk) for the unknown variance case. A year later, Owen
(1967), developed a variable sampling plan for two specification levels. Variables-acceptance
sampling plans were covered in many books. Among these are: Bowker and Goode (1952), Duncan
(1974), Guenther (1977). Grant and Leavenworth (1980), and Schilling (1982).
Plans for non-normal distributions have been considered in the literature, as long as the form of the
underlying distribution is known and the proportion of defectives could be calculated just from
knowing the parameters of the distribution, usually the mean. Srivastava (1961) and Mitra and Das
(1964) investigated the effect of non-normality for one-sided plans. Zimmer and Burr (1963)
developed variables sampling plans using measures of skewness and kurtosis to adjust for nonnormality. Guenther (1972) introduced variables sampling plans when the underlying distribution is
Poisson or binomial. Duncan (1974), in his book, developed a plan under the assumption that the
distribution of the observations is a member of the Type III Pearson family, which in its standardized
form depends only on the coefficient of skewness. Guenther (1977), in his book, considered the
exponential family as the underlying distribution. Lam (1994) developed Bayesian variables sampling
plans for the exponential distribution with Type I censoring. Suresh and Ramanathan (1997) derived a
plan for symmetric underlying distributions. Plans for autocorrelated observations are described in
Hapuarachchi and Macpherson (1992).
In this article, the variables acceptance approach is used to study the problem of listing sites under
the Clean Water Act. Methods based on use of actual values are shown to have error rates superior to
methods based on binary information. Positive autoconelation in data is a potential problem and is
shown to increase the Type I error rates. Two methods for adjusting for autocorrelation are presented
and error rates are calculated for these methods. Estimates of sample size are given that will achieve
specific error rates for independent and autocorrelated situations. The approach taken is similar to that
in Barnett and Bown (2002), who develop a testing approach for composite samples.
2. VARIABLES PLAN FOR THE NORMAL DISTRIBUTION
A variety of tests are discussed in the acceptance sampling literature for normally distributed data. All
involve calculation of a critical value for a test based on the mean and are summarized in Table 1.
Owen (1966,1967) (see also Duncan, 1974) presented the acceptance sampling by variables technique
when the distribution of the measurements is normal with either known or unknown standard
deviation. This section summarizes that material in the context of environmental sampling.
Copyright 0 Z W 3 John Wiley &Sons. Ltd.
Envimnmerrics 2003: 14: 373-386
377
EVALUATION OF WATER QUALITY
Table 1. Cutoffs for different methods of testing Ho :p 5 po versus HI:p > po using the test statistic for a
lower standard L of the form test = (Z. - L)/s or a test slatistic of the form test = (U -%)Is for an upper
standard. The value in the table is the cutoff k that defines the rejection region. For example, with a lower
standard we reject if lest = (.Z - L)/s < k
Method
Plotting acronym
Cutoff
b
k =--
Normal, variance known
\h
ZPJ
iot/2ni$
Wallis approximation
Wallis
k=
Normal, variance unknown (non-central 1)
NCT
k=-
+ 4n - 222, - 2 n z ~
2n - i?
J;; where X = -J;izPO
AR(1) using normal approximation (Hapuardchchi plan)
HAP
ARO) using non-central t
whete X =
-dB; Suppose that the environmental measurements have a normal distribution with mean p and standard
deviation u. The criterion is assumed to be some upper (lower) specified limit U(L). Then the
- U ) / u ] ,where is the cumulative
proportion of defective items p can be calculated aspu = a[(@
distribution function for a standard normal random variable (analogously p~ = + [ ( L - p ) / u ] ) .
Specifying p , u and U will specify p. Moreover, if u is constant, then p will depend only on U and
p. In this case, instead of testing hypotheses about p we can rather evaluate hypotheses about p and
then base our analysis on the sample mean E. Therefore, to test Ho : p 5 po (no impairment, don't list)
versus HI: p > po (impairment, list), we can proceed in two ways. First, if there is a single upper limit
U, use the sample mean to compute zu = (U - E ) / u and then reject Ho if z u c k , where
k = z,/&
- b o . In the second procedure, the proportion of impaired samples is estimated by
and the null hypothesis is rejected i f j , > m, where m = ~ ( k - ) .
It is notable that the second
procedure may be preferable since it gives adecision based on the probability of a defective item. If we
-i
L)/u and k will he the same. The
have a lower standard instead of an upper one, then z~ = (
estimated proponion of impaired sites is
Copyright
02003 John Wiley & Sons, Lrd.
378
E.P. SMITH ETAL.
Owen also presented an analysis for the case of double specification limits, i.e. lower and upper
limits.
2.1. Normal distribution with unknown variance
Lieberman and Resnikoff (1955), developed acceptance sampling plans for data from a normal
distribution with unknown variance. A reasonable test statistic for the hypothesis using a lower limit is
which has a non-central t distribution with n - 1 degrees of freedom and non-centrality parameter
X = -zp&. The hypothesis is rejected if t < ~,,-I,A,*.
The power curve could be obtained by
calculating Pr(t < &k) = PI(! < t n - l , ~ , l - ~ )for specified values of p or p and Type I1 error p.
Here k = t . - l , ~ , ~ / &An
. alternative is to use
The hypothesis is rejected if t < k.
Wallis (1947) suggested an alternative approach. This approach is approximate and was designed to
avoid the use of the non-central t distribution. The test statistic is still the same and the value of k is
given by
k=
z, 42-
- 2nzN
2n - zi
(3)
2.2. Tests with autocorrelated data
A potentially important problem in using measurements collected over time is the possibility that the
measurements are correlated. Hapuarachchi and Macpherson (1992) studied the effect of serial
corielation on acceptance sampling plans by variables assuming the measurements follow an
autoregressive of order p [AR(p)] process. They assumed that the flh measurement x, is modeled as
with p a constant ( usually the mean of the process) and the e,'s assumed to be independently normally
distributed with mean of zero and variance of
In the model (4), y, is said to follow an AR(p)
process. A special case is the AR(1) process specified by
4.
This model is more common to water quality issues.
For an AR(1) process it can be shown that the approximate sample variance of the mean is (Darken
et al,, 2000)
Copyright O 2003 john Wiley & Sms, Ltd.
EVALUATION OF WATER QUALITY
379
An approximate test for HO: p 2 /& versus HI: p < p~ would again be based on the t statistic. The
critical value k would be based on a non-central r distribution with n - 1 degrees of freedom and
non-centrality parameter X = -z,\l[n(l - 6)]/(1 6), where 6 is the estimated autocorrelation. So,
we are using the same test statistic when assuming independence; however, the cutoff point of the test
is modified to take the dependency among the measurements into account. Note that the distribution of
the test statistic depends on the autoregressive parameter 9.
Also note that, with a positive autocorrelation structure, the non-centrality parameter and hence the
cutoff point is less than the independence cutoff point. Thus, rejection is not as easy as it is in the
independence case.
+
3 . SAMPLE SIZE
One use of the above methodology is to set up a sampling plan that attempts to control the error rates.
In particular suppose that it is desired to set up a variables plan which has a probability a of falsely
declaring a site impaired for a site with an acceptable proportion po of measurements actually more
extreme than the standard and a probability for falsely not listing an impaired site that has an
unacceptable proportion pl of measurements more extreme than the standard. Typically, pl is a value
chosen to be a worst case in the sense that we want a site listed with high probability if this proportion
of measurements does not meet the water quatity standard. Under the first procedure (normal
distribution with known variance) with an upper standard we will not declare impairment if
Z, = [(U - .t)/u]2 k. It can be shown (Duncan, 1974) that the suitable sample size n is given by
where z,,zp,zp,
distribution.
and z,, represent respectively the a, P, po and pl quantiles of the standard normal
3.1. Sample size with unknown vuriance
When the standard deviation is unknown, almost the same procedure for the known standard deviation
is used to calculate sample size except that we replace the population standard deviation u by the
sample standard deviation s. Three methods are available to derive a variables plan in this case. In
the first method, special formulas for finding n and k, according to certain values of po, p , , a and p,
are introduced by Wallis (1947). They are
and
These formulas are based on the assumption that T.
k2/2n).
mean of p f k u and variance $ ( l / n
+
Copyright O 2003 John Wiley & Sons, Ltd.
f ks is approximately normally distributed with a
Environn~etrics2003; 14: 373-386
380
E. P. SMITH ETAL.
Table 2. Estimated lower values (i)and cutoffs for test for impairment based on different statistics and Type I
error rates for dissolved oxygen data. If the lower value is smaller than 5.0, this indicates listing the site. The
autocorrelation was estimated to be 0.6699 and the calculated test statistic is 1.135
Methad
Type I error rate
0.05
I
Normal
AR(I)
0.891
0.885
0.522
0.972
0.965
0.687
0.10
t
Estimated lower value (i)
Cutoff (k)
Normal
ARI.I ).
5.44
5.45
6.10
5.29
5.31
5.80
I
Normal
ARf 1)
The second method is from a monograph due to L. J. Jacobson. Using the Jacobson monograph and
specified values of po,pl, a and 0, we can calculate n and k. The Jacobson approach is presented in
Duncan (1974, Ch. 12, p. 271). However, the results obtained by this method are not precise and Owen
mentioned a third method that gives more precise results.
In the third method, the analysis is based on the exact distribution rather than the approximate
distribution proposed by Wallis. The exact distribution when using the sample standard deviation
instead of the population standard deviation is the non-central t distribution with n - 1 degrees of
freedom. To find n and k i n this case, a special table specifically designed for this purpose has been
prepared by Owen (1963, Table 2). Lieberman and Resnikoff (1955) provide special tables to get jj.
and m according to a given sample size n and accepted quality level (i.e. p a ) These tables are found in
(Lieberman and Resnikoff, 1955). Another set of tables was produced in Statistical Research Group
(1947, pp. 22-25).
Hapuarachchi and Macpherson (1992) describe sev&al approaches for estimating sample size when
data are autocotrelated. In the case of unknown variance, the authors note that the distributions of
needed quantities are difficult to obtain. For the AR(1) case, they suggest using an approximate sample
size given by
Table 3. Sample sizes calculated to achieve pre-determined ermr rates for different methods with
autocorrelation set to zero. Values were rounded UD to the nearest integer
-
"w 1
0.05
0.1
-
Type 11
Normal (u known)
Wallis approx.
Non-central t
0.05
0.1
0.2
0.05
30
24
17
24
58
44
30
48
44
34
24
Copyright O 2W3 John Wiley 8r Sons, Ltd
76
EVALUATION OF WATER QUALITY
Table 4. Sample sizes calculated to achieve pre-determined error rates for
different methods with varying autocorrelation. Values were rounded up to the
nearest integer. The Tyve I and II ermr rates were both set to 0.2
Autocorrelation
HAP
NCT AR(I)
which is just the normal based approach (Equation 7) adjusted for antocorrelation. One might also adapt
the sample sizes calculated using one of the other methods for the correlation by multiplying by a factor
[(I @/(I - g)]. Another approach would use the non-central t distribution with the adjustment.
+
4. COMPARISON OF METHODS
Given the variety of methods a comparison is warranted. It is possible to analytically compare some of
the methods under certain conditions. In situations where the error rates could be exactly calculated for
all methods (Figures 2 and 3) we used the analytical approach. When analytical solutions were not
possible for all the methods we relied on a simulation for the comparisons. In, the simulations,
normally distributed data were generated with different levels of autocorrelation (B=0.0, 0.25, 0.5,
TEST
- b~nom~al
---
0.0
0
10
20
30
40
50
60
NC T
raw score
- - - Wallis
Sample size
-
-
-
Figure 3. Plot of vrobabilitv of not listine an im~airedsite (Bi
... "sine binomial. raw score and acceotance sarn~line
. methods.
Note that the probability an unimpaired site is listed (a)is set to 0.2 for the methods. This e m rate is not set for the raw score
method. Values of thiserror ratearegiven in Figure2. The parametersofthe normaldistribution were set so the vrobabilitv that a
measurement was more extreme than the standard. The binomial probability of impairment was set to 0.25
Copyright 0 2003 John Wiley & Sons. Ltd.
~~~
~~
E,rvimnn~etrics2003: 14: 373-386
382
E. P. SMITHETAL.
0.75) and evaluated using cutoff points associated with different Type I error rates (a= 0.05,0.1,0.2).
For estimation of Type I error rates the site was assumed to meet standards (po = 0.1), while for the
estimation of Type I1 error rates the site was assumed to not meet standards (p, = 0.25). For
estimation of the e m r rate for a given case, 10000 times series were generated with sample sizes
ranging from 5 to 50. A 'bum-in' of 20 points was used to stabilize the time series (e.g. for the
case N s 5 . 2 5 observations were generated but only the last 5 were used). For purposes of space just
the results for the cases a = 0.2 and 0 = 0.0 and 0.5 are discussed.
Figure 2 displays the V p e I error rates for the binomial and raw score methods with no
autocorrelation. What is obvious from this display is that the binomial method bounds the Type 1
error by choice of cutoff to be below 0.20. The raw score method results in a rather large Type I error.
The raw score method may be viewed as a binomial method with changing Type I error rate.
Alternatively the binomial method may be viewed as a raw score method with varying cut point.
The normal based methods with no autocorrelation are not presented as the error rate is set to a fixed
value.
Figure 3 displays the Type 11error rate for the two binary methods, the non-central t approach and
Wallis' approach with at Type I error set at 0.1. The normal case with known variance was not
considered. The Type I1 error rate is large for the binomial method relative to the raw score. This result
makes sense since the raw score has a high Type I error rate and hence a low cutoff relative to the
binomial. Therefore it will be more powerful. The normal-based approaches using the actual data
produce Type I1 error rates similar to the raw score method. Thus if the data are consistent with the
normal assumption, there is much to gain from using the normal based methods as the Type I error rate
can be fixed. The method is consistent with the low Type I e m r rates for the binomial and has similar
Type 11 error rates to the raw score. We note that there is little differencein terms of error rates for the
non-central t and Wallis' method.
Figures 4 and 5 plots this error rate for the tests when the autocorrelation is 0.0 and 0.5. When the
correlation is zero (Figure 4), we find that the Type I error rates are as expected with the exception of
the approximation proposed by Hapuarachchi and Macpherson. For this method, the error rate is
higher than expected. The inflation in the error rate is due to the use of the estimated standard deviation
rather than the actual standard deviation in the test statistic. Type I1 error rates for the different
methods based on the normal model yield better error rates than either the binomial or raw score
approach. These results indicate that if the assumption of normality is met, there is considerable
advantage to using a parametric method.
When the autocorrelation is 0.5, we find that not accounting for positive autocorrelation results in
an increase in the Type I error rate (Figure 5) for most of the procedures. The error rates for the noncentral t, Wallis and binomial methods exceed the preset level of 0.2. The error rate for the non-central
t is close to the nominal value for small sample sizes but increases with sample size. Error rates for the
autocorrelation-adjusted tests are less than the desired level for small sample sizes but approach the
level as the sample size increases. Also note that the maximum error rate for the raw score method is
slightly decreased over that of Figure 4. When Type I1 error rates are considered, we find that the
autocorrelation increases the error rate relative to the uncorrelated case (compare Figures 4b and 5b).
The Type 11 error rates for the AR(1) non-central r and the method of Hapuarachchi and Macpherson
are higher than those of the non-central t and Wallis' method. Relative to the binomial, the
autocorrelation-adjusted tests generally have smaller error rates.
Also apparent from the figures is that the Type 11 error rates will be large for small sample sizes.
When the data are not correlated, samples of size 10 might yield reasonable results if one would be
willing to balance Type I and Type I1 error rates at around 0.2. In the case of correlated data, additional
Copyright 0 2003 John Wiley &Sons, Ltd.
Environmetrics 2003: 14: 373-386
EVALUATION OF WATER QUALITY
a
$
0.6
-c Bin
0.5
w
'0.4
-
+Wallis
5 0.2
0.1
E=O
0
",%9+,,!.,,9,%49@9
Sample slre
-2
-i
8,
w
=0
r-gi
0.6
0.5
0.4
0.3
+Wdllis
1: 0.2
F 0.1
E=o
0
6 6 6 9 + + 3 % 4 9 @ 4
Sample alze
Figure 4. (a) Simulated actual 5 p e I error rates far different methods using u = 0.2 and autocorrelation of0.0 (b) Type 11 error
rates using 0 = 0.2
samples would be required to produce the same error rates and a sample size of 20 is better although
the number of additional samples depends on the size of the autocorrelation.
5. EXAMPLE
As an example we consider monthly data collected over a two-year period on dissolved oxygen. The
values are plotted in Figure 6. Interest is achieving a (lower) standard of 5.0 withpo = 0.1. The sample
mean is 1 = 7.03 with standard deviations = 1.787, n = 24 and a = 0.05. The calculated value of the
test statistic is 1.135. Using a non-central t distribution we find the critical value k=0.890. So, one
does not reject the null hypothesis at a=0.05 and the site would not be listed. Assuming an AR(1)
process, the autocorrelation coefficient is calculated as 0.6699. We find k = 0.521. Using the adjusted
non-central t statistics, the decision would be to not reject the null hypothesis at a=0.05. An
alternative approach is to base results on the estimated limit. Table 2 provides a summary of three
different tests at different a levels. Id all cases, the estimated lower value (i)is above 5.0 so one would
not reject.
Copyright O 2003 John Wiley & Sons. Ltd.
Euvironrr~crrics2003; 14: 373-386
E.P.SMITH ETAL.
k. 0.5
+Bin
d
-5 0.4
0.3
-g; 0.2
U
+Wailis
I-* 0.1
€=0.5
0
* + Q + ' + + 5 $ $ 9 $ @
Sample size
0.6
1
+Bin
0.5
- 0.3
0
e: 0.2
-rcWallis
F 0.1
e0.5
0
% & 8 + ' + @ . - j $ + 9 $ @
Sample sire
Figure 5. (a) SimulatedV p e I error rates using ol=0.2 for autoconelated data with correlation of 0.5; (b) simulated V p e II
enor rates using a=0.2 for autocarreiated data with correlation of 0.5
. . . . . . . . . . . . . . . . . . . . . . . .
04
1995
1996
Sampllng Date
Figure 6. Plot of 24 observations on dissolved oxygen for a two-year period from 1995 to 1997
Copyright O 2003 John Wiley & Sans. Ltd.
Envimnmelrics 2003; 14: 373-386
EVALUATION OF WATER QUALITY
385
Sample size calculations are presented in Tables 3 and4. Table 3 provides different sample sizes for
varying Types I and I1 error rates with autocorrelation set to zero. For most cases, sample sizes are
greater than those achieved by quarterly sampling in a two-year period. Unless error rates are set to be
large (0.2), relatively large sample sizes are required. Table 4 gives estimates of sample sizes required
under autocorrelation with Types I and I1 error rates set to 0.2. As indicated in the table, as the
autocorrelation increases, so does the required sample size. With correlations that might be expected in
water studies (0.5), roughly 6 years of quarterly data would be required.
6. DISCUSSION
The investigation of error rates reveals that the methods currently in use may be improved by adopting
approaches based on acceptance sampling by variables. In particular, methods based on the noncentral t distribution produce error rates better than the methods based on discretized data. When
measurements are made over time correlation is to be expected and adjustments to the test are
available. Positive autocorrelation effectively reduces the sample size. Current protocols use around
eight observations over a two-year period to make inferences. Based on the analysis of error rates and
dissolved oxygen data, the sample size seems inadequate to make strong inferences, even if error rates
around 0.2 are desired.
Our approach is based on testing and the use of methods from acceptance sampling. Another
approach may be based on the use of tolerance intervals (Gibbons, 1994; Hahn and Meeker, 1991). The
approach based on tolerance intervals using a non-central t distribution is equivalent to the acceptance
sampling approach using the non-central t. The tolerance limit approach was suggested by Lin et a!.
(2000) for use with 303(d) listing. Smith (2002) describes a generalized approach inwhich samples
from different times or spatial locations are used with a tolerance interval for setting standards.
Many of decisions made about water quality are based on small sample sizes. For example, if it is
required to produce a report every two years, the evaluation of water quality might be based on
quarterly sampling, yielding eight observations. In cases with small sample sizes decisions may be
affected by variation due to estimation of variance. One approach for improving estimation of variance
is to use data from alternate sources. For example, there may be several locations within a watershed.
The information from the multiple sites may be pooled to produce a variance estimate. Estimation of
variance may also use a Bayesian approach.
The above literature does not cover multivariate data sets. There is not much work done for the
multivariate setup. The first work published on this problem was of Baillie (1987a,b). He developed
direct multivariate generalizations of the variables-acceptance sampling procedures under the
assumption of multivariate normal distribution. A few years later, Hamilton and Lesperance (1991)
dropped the assumption of multivariate-normality and proposed a method to deal with multivariate
data using Wallis' (1947) advice. They transformed the bivariate data into one normal variable, and
then applied the univariate techniques to their new variable.
ACKNOWLEDGEMENTS
This research has been supported through grants from the Virginia Water Resources Research Center and U.S.
Environmental ~rotectionAgency's ~ccnc; to Achieve ~ e s u 6 s(STAR) (grant no. R82795301). Although the
research described in the article has been funded wholly or in part by the U.S.Environmental Protection Agency's
STAR proarams, it has not been subiected to any EPA review and therefore does not necessarilv reflect the views
of the kggncy, and no official endoisement should be inferred.
Copyright O 2033 John Wiley & Sons. Ltd.
Envimnmerrics 2003; 14: 373-386
386
E. P. SMITH ETAL.
REFERENCES
Baillie DH. 1987a. Multivariate acceptance sampling: some applications to defense procurement. Slatislicion 36: 465478.
Baillie DH. 1987b. Multivariate acceptance sampling. In Fmntiers in Slalislical Quality Conlml 3. Lenz HJ, Wetherill GG.
Wilrich
(eds). Physica Verlag: WUerzbirg, Vienna.
Bamett V, Bown M. 2002. Statisticaliy meaningful standards facontaminated sites using composite sampling. Envimnmetrier
13: 1-14.
Rowker
AH. Goode
HP. 1952. Samoline I"$~ectionbv Variables. McGraw-Hill: New York.
-~
~~~~~~~.
-~~~~~
Darken PF, Holzman GI, Smith EP, Z i p p r CE. 2000. Detecting changes in trends in water quality using modified Kendall's tau.
Envimnmelries
11: 423-434.
~.~
Duncan AJ. 1974. Quality Contml and Indurrriai Slotisrics. Iwin: Homewood, IL.
Gibbons RD. 1994. Sl~risricolMethodsfor Gmundwaler Monitoring. John Wiley & Sons: New York.
Grant EL. Leavenworth RS. 1980. Statistical Quality Contml, 5th edition. McGraw-Hill: New York.
Guenther WC. 1972. Variables sampling plans for the Poisson and Binomial. Stotistica Neerlandico 26: 17-24.
Guenther WC. 1977. Sampling Inspection in Statistical Quality Conlml. Macmillan Publishing Co., Inc: New York.
Hshn GJ, Meeker WQ. 1991. Slallstieal Inlervols: A Guidefor Pmetitioners. John Wiley & Sons: New York.
Hamilton DC, LesperanceML. 1991. A consultingproblem involving bivariate acceptance sampling by variables. The Canadian
Journal of Storistics 19: IOP-117.
Hapuarachchi KP, Macpherson BD. 1992. Autoregressive processes applied to acceptance sampling by variables. Communications in Slotistics: Simulation 21(3): 833-848.
Lam Y. 1994. Bayesian variables sampling plans for the exponential distribution with Qpe I censoring. Anttals of Statistics 22:
~
~
~
-,"-...,
LO&-71 1
Lieberman GI, Resnikoff GI. 1955. Sampling
. -~plans for inspection by variables. Journal of ths American Stolisticai Association
50: 457-516.
Lin P. Meeter D. Niu X. 2000. A nonparameuic procedure for listing and delisting impaired waters based on criterion
exeeedences. Technical Repon submitted to the Florida Department of Environmental Protection,
McBride GB. Ellis IC. 2001. Confidence of compliance: a Bayesian approach for percentile standards. Water Research 35:
I I n-I124.
Mitra SK, Das NO. 1964. The effect of "on-normalily on sampling inspection. Sankhyr?26(A): 169-176.
Owen DB. 1963. Faelorsfor one-sided tolemnce limits and for variables sampling plans. Monograph No. SCR-607, Sandia
Corporation.
Owen DB. 1966. One-sided variables sampling plans. Indusrriol Quolily Conrml22: 4 5 W 5 6 .
Owen DB. 1967. Variables sampling plans based on the normal distribution. Technometries 9: 417423.
Schilling EG. 1982. Acceptance Sampling in Quality Conlml. Marcel Dekker: New York.
Smith EP, Ye K. Hughes C. Shabman L. 2WI. Statistical assessment of violations of water quality standards under Section
303(d) of the Clean Water Act. Envimnmental Science and Technology 35: 606412.
Smith R. 2002. The use of random-model tolerance intervals in environmental monitoring and regulation. Journal of
Agn'culturol, Bioiogicoi. and Envimnmenloi Starislics 7: 74-94.
Srivasldva ABL. 1961. Variables sampling inspection for non-normal samples. Journal of Science and Engineering Reseorch 5:
145-152.
Statistical Research Group, Columbia University. 1947. Techniques of Slolislicol Analysis. McGraw-Hill Book Co., Ine: New
Vnrk
Suresh RP. Ramannthan TV. 1997. Acceptance sampling plans by variables for a class of symmetric distributions. Communications in Statistics: Simuiotion 26: 1379-1391.
Wallis WA. 1947. Use of variables in acceptance inspection for percent defective. In Seieeled Techniques of Statistical Anolysis
for Scienr$c and Industrial Research and Production and Management Engineering. McGraw Hill: New York; 3-93.
Zimmer WJ, Burr IW. 1963. Variables sampling plans based on non-normel papulatians. hdustrial Quality Contml21: 18-26.
Copyright O 2003 John Wiley &Sons, Ltd.